Actual source code: blzpack.c
1: /*
2: This file implements a wrapper to the BLZPACK package
4: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
5: SLEPc - Scalable Library for Eigenvalue Problem Computations
6: Copyright (c) 2002-2011, Universitat Politecnica de Valencia, Spain
8: This file is part of SLEPc.
9:
10: SLEPc is free software: you can redistribute it and/or modify it under the
11: terms of version 3 of the GNU Lesser General Public License as published by
12: the Free Software Foundation.
14: SLEPc is distributed in the hope that it will be useful, but WITHOUT ANY
15: WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
16: FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for
17: more details.
19: You should have received a copy of the GNU Lesser General Public License
20: along with SLEPc. If not, see <http://www.gnu.org/licenses/>.
21: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
22: */
24: #include <private/epsimpl.h> /*I "slepceps.h" I*/
25: #include <../src/eps/impls/external/blzpack/blzpackp.h>
27: PetscErrorCode EPSSolve_BLZPACK(EPS);
29: const char* blzpack_error[33] = {
30: "",
31: "illegal data, LFLAG ",
32: "illegal data, dimension of (U), (V), (X) ",
33: "illegal data, leading dimension of (U), (V), (X) ",
34: "illegal data, leading dimension of (EIG) ",
35: "illegal data, number of required eigenpairs ",
36: "illegal data, Lanczos algorithm block size ",
37: "illegal data, maximum number of steps ",
38: "illegal data, number of starting vectors ",
39: "illegal data, number of eigenpairs provided ",
40: "illegal data, problem type flag ",
41: "illegal data, spectrum slicing flag ",
42: "illegal data, eigenvectors purification flag ",
43: "illegal data, level of output ",
44: "illegal data, output file unit ",
45: "illegal data, LCOMM (MPI or PVM) ",
46: "illegal data, dimension of ISTOR ",
47: "illegal data, convergence threshold ",
48: "illegal data, dimension of RSTOR ",
49: "illegal data on at least one PE ",
50: "ISTOR(3:14) must be equal on all PEs ",
51: "RSTOR(1:3) must be equal on all PEs ",
52: "not enough space in ISTOR to start eigensolution ",
53: "not enough space in RSTOR to start eigensolution ",
54: "illegal data, number of negative eigenvalues ",
55: "illegal data, entries of V ",
56: "illegal data, entries of X ",
57: "failure in computational subinterval ",
58: "file I/O error, blzpack.__.BQ ",
59: "file I/O error, blzpack.__.BX ",
60: "file I/O error, blzpack.__.Q ",
61: "file I/O error, blzpack.__.X ",
62: "parallel interface error "
63: };
67: PetscErrorCode EPSSetUp_BLZPACK(EPS eps)
68: {
70: PetscInt listor,lrstor,ncuv,k1,k2,k3,k4;
71: EPS_BLZPACK *blz = (EPS_BLZPACK *)eps->data;
72: PetscBool issinv;
75: if (eps->ncv) {
76: if (eps->ncv < PetscMin(eps->nev+10,eps->nev*2))
77: SETERRQ(((PetscObject)eps)->comm,0,"Warning: BLZpack recommends that ncv be larger than min(nev+10,nev*2)");
78: }
79: else eps->ncv = PetscMin(eps->nev+10,eps->nev*2);
80: if (eps->mpd) { PetscInfo(eps,"Warning: parameter mpd ignored\n"); }
81: if (!eps->max_it) eps->max_it = PetscMax(1000,eps->n);
83: if (!eps->ishermitian)
84: SETERRQ(((PetscObject)eps)->comm,PETSC_ERR_SUP,"Requested method is only available for Hermitian problems");
85: if (eps->which==EPS_ALL) {
86: if (eps->inta==0.0 && eps->intb==0.0) SETERRQ(((PetscObject)eps)->comm,1,"Must define a computational interval when using EPS_ALL");
87: blz->slice = 1;
88: }
89: PetscTypeCompare((PetscObject)eps->OP,STSINVERT,&issinv);
90: if (blz->slice || eps->isgeneralized) {
91: if (!issinv) SETERRQ(((PetscObject)eps)->comm,PETSC_ERR_SUP,"Shift-and-invert ST is needed for generalized problems or spectrum slicing");
92: }
93: if (blz->slice) {
94: if (eps->intb >= PETSC_MAX_REAL) { /* right-open interval */
95: if (eps->inta <= PETSC_MIN_REAL) SETERRQ(((PetscObject)eps)->comm,1,"The defined computational interval should have at least one of their sides bounded");
96: STSetDefaultShift(eps->OP,eps->inta);
97: }
98: else { STSetDefaultShift(eps->OP,eps->intb); }
99: }
100: if (!eps->which) {
101: if (issinv) eps->which = EPS_TARGET_REAL;
102: else eps->which = EPS_SMALLEST_REAL;
103: }
104: if ((issinv && eps->which!=EPS_TARGET_REAL && eps->which!=EPS_TARGET_MAGNITUDE && eps->which!=EPS_ALL) || (!issinv && eps->which!=EPS_SMALLEST_REAL))
105: SETERRQ(((PetscObject)eps)->comm,1,"Wrong value of eps->which");
107: k1 = PetscMin(eps->n,180);
108: k2 = blz->block_size;
109: k4 = PetscMin(eps->ncv,eps->n);
110: k3 = 484+k1*(13+k1*2+k2+PetscMax(18,k2+2))+k2*k2*3+k4*2;
112: listor = 123+k1*12;
113: PetscFree(blz->istor);
114: PetscMalloc((17+listor)*sizeof(PetscBLASInt),&blz->istor);
115: blz->istor[14] = PetscBLASIntCast(listor);
117: if (blz->slice) lrstor = eps->nloc*(k2*4+k1*2+k4)+k3;
118: else lrstor = eps->nloc*(k2*4+k1)+k3;
119: lrstor*=10;
120: PetscFree(blz->rstor);
121: PetscMalloc((4+lrstor)*sizeof(PetscReal),&blz->rstor);
122: blz->rstor[3] = lrstor;
124: ncuv = PetscMax(3,blz->block_size);
125: PetscFree(blz->u);
126: PetscMalloc(ncuv*eps->nloc*sizeof(PetscScalar),&blz->u);
127: PetscFree(blz->v);
128: PetscMalloc(ncuv*eps->nloc*sizeof(PetscScalar),&blz->v);
130: PetscFree(blz->eig);
131: PetscMalloc(2*eps->ncv*sizeof(PetscReal),&blz->eig);
133: if (eps->extraction) { PetscInfo(eps,"Warning: extraction type ignored\n"); }
135: EPSAllocateSolution(eps);
137: /* dispatch solve method */
138: if (eps->leftvecs) SETERRQ(((PetscObject)eps)->comm,PETSC_ERR_SUP,"Left vectors not supported in this solver");
139: eps->ops->solve = EPSSolve_BLZPACK;
140: return(0);
141: }
145: PetscErrorCode EPSSolve_BLZPACK(EPS eps)
146: {
148: EPS_BLZPACK *blz = (EPS_BLZPACK *)eps->data;
149: PetscInt nn;
150: PetscBLASInt i,nneig,lflag,nvopu;
151: Vec x,y;
152: PetscScalar sigma,*pV;
153: Mat A;
154: KSP ksp;
155: PC pc;
156:
158: VecCreateMPIWithArray(((PetscObject)eps)->comm,eps->nloc,PETSC_DECIDE,PETSC_NULL,&x);
159: VecCreateMPIWithArray(((PetscObject)eps)->comm,eps->nloc,PETSC_DECIDE,PETSC_NULL,&y);
160: VecGetArray(eps->V[0],&pV);
161:
162: if (eps->isgeneralized && !blz->slice) {
163: STGetShift(eps->OP,&sigma); /* shift of origin */
164: blz->rstor[0] = sigma; /* lower limit of eigenvalue interval */
165: blz->rstor[1] = sigma; /* upper limit of eigenvalue interval */
166: } else {
167: sigma = 0.0;
168: blz->rstor[0] = eps->inta; /* lower limit of eigenvalue interval */
169: blz->rstor[1] = eps->intb; /* upper limit of eigenvalue interval */
170: }
171: nneig = 0; /* no. of eigs less than sigma */
173: blz->istor[0] = PetscBLASIntCast(eps->nloc); /* number of rows of U, V, X*/
174: blz->istor[1] = PetscBLASIntCast(eps->nloc); /* leading dimension of U, V, X */
175: blz->istor[2] = PetscBLASIntCast(eps->nev); /* number of required eigenpairs */
176: blz->istor[3] = PetscBLASIntCast(eps->ncv); /* number of working eigenpairs */
177: blz->istor[4] = blz->block_size; /* number of vectors in a block */
178: blz->istor[5] = blz->nsteps; /* maximun number of steps per run */
179: blz->istor[6] = 1; /* number of starting vectors as input */
180: blz->istor[7] = 0; /* number of eigenpairs given as input */
181: blz->istor[8] = (blz->slice || eps->isgeneralized) ? 1 : 0; /* problem type */
182: blz->istor[9] = blz->slice; /* spectrum slicing */
183: blz->istor[10] = eps->isgeneralized ? 1 : 0; /* solutions refinement (purify) */
184: blz->istor[11] = 0; /* level of printing */
185: blz->istor[12] = 6; /* file unit for output */
186: blz->istor[13] = PetscBLASIntCast(MPI_Comm_c2f(((PetscObject)eps)->comm)); /* communicator */
188: blz->rstor[2] = eps->tol; /* threshold for convergence */
190: lflag = 0; /* reverse communication interface flag */
192: do {
193: BLZpack_(blz->istor,blz->rstor,&sigma,&nneig,blz->u,blz->v,&lflag,&nvopu,blz->eig,pV);
195: switch (lflag) {
196: case 1:
197: /* compute v = OP u */
198: for (i=0;i<nvopu;i++) {
199: VecPlaceArray(x,blz->u+i*eps->nloc);
200: VecPlaceArray(y,blz->v+i*eps->nloc);
201: if (blz->slice || eps->isgeneralized) {
202: STAssociatedKSPSolve(eps->OP,x,y);
203: } else {
204: STApply(eps->OP,x,y);
205: }
206: IPOrthogonalize(eps->ip,0,PETSC_NULL,eps->nds,PETSC_NULL,eps->DS,y,PETSC_NULL,PETSC_NULL,PETSC_NULL);
207: VecResetArray(x);
208: VecResetArray(y);
209: }
210: /* monitor */
211: eps->nconv = BLZistorr_(blz->istor,"NTEIG",5);
212: EPSMonitor(eps,eps->its,eps->nconv,
213: blz->rstor+BLZistorr_(blz->istor,"IRITZ",5),
214: eps->eigi,
215: blz->rstor+BLZistorr_(blz->istor,"IRITZ",5)+BLZistorr_(blz->istor,"JT",2),
216: BLZistorr_(blz->istor,"NRITZ",5));
217: eps->its = eps->its + 1;
218: if (eps->its >= eps->max_it || eps->nconv >= eps->nev) lflag = 5;
219: break;
220: case 2:
221: /* compute v = B u */
222: for (i=0;i<nvopu;i++) {
223: VecPlaceArray(x,blz->u+i*eps->nloc);
224: VecPlaceArray(y,blz->v+i*eps->nloc);
225: IPApplyMatrix(eps->ip,x,y);
226: VecResetArray(x);
227: VecResetArray(y);
228: }
229: break;
230: case 3:
231: /* update shift */
232: PetscInfo1(eps,"Factorization update (sigma=%g)\n",sigma);
233: STSetShift(eps->OP,sigma);
234: STGetKSP(eps->OP,&ksp);
235: KSPGetPC(ksp,&pc);
236: PCFactorGetMatrix(pc,&A);
237: MatGetInertia(A,&nn,PETSC_NULL,PETSC_NULL);
238: nneig = PetscBLASIntCast(nn);
239: break;
240: case 4:
241: /* copy the initial vector */
242: VecPlaceArray(x,blz->v);
243: EPSGetStartVector(eps,0,x,PETSC_NULL);
244: VecResetArray(x);
245: break;
246: }
247:
248: } while (lflag > 0);
250: VecRestoreArray(eps->V[0],&pV);
252: eps->nconv = BLZistorr_(blz->istor,"NTEIG",5);
253: eps->reason = EPS_CONVERGED_TOL;
255: for (i=0;i<eps->nconv;i++) {
256: eps->eigr[i]=blz->eig[i];
257: }
259: if (lflag!=0) {
260: char msg[2048] = "";
261: for (i = 0; i < 33; i++) {
262: if (blz->istor[15] & (1 << i)) PetscStrcat(msg,blzpack_error[i]);
263: }
264: SETERRQ2(((PetscObject)eps)->comm,PETSC_ERR_LIB,"Error in BLZPACK (code=%d): '%s'",blz->istor[15],msg);
265: }
266: VecDestroy(&x);
267: VecDestroy(&y);
268: return(0);
269: }
273: PetscErrorCode EPSBackTransform_BLZPACK(EPS eps)
274: {
276: EPS_BLZPACK *blz = (EPS_BLZPACK *)eps->data;
279: if (!blz->slice && !eps->isgeneralized) {
280: EPSBackTransform_Default(eps);
281: }
282: return(0);
283: }
287: PetscErrorCode EPSReset_BLZPACK(EPS eps)
288: {
290: EPS_BLZPACK *blz = (EPS_BLZPACK *)eps->data;
293: PetscFree(blz->istor);
294: PetscFree(blz->rstor);
295: PetscFree(blz->u);
296: PetscFree(blz->v);
297: PetscFree(blz->eig);
298: EPSFreeSolution(eps);
299: return(0);
300: }
304: PetscErrorCode EPSDestroy_BLZPACK(EPS eps)
305: {
309: PetscFree(eps->data);
310: PetscObjectComposeFunctionDynamic((PetscObject)eps,"EPSBlzpackSetBlockSize_C","",PETSC_NULL);
311: PetscObjectComposeFunctionDynamic((PetscObject)eps,"EPSBlzpackSetNSteps_C","",PETSC_NULL);
312: return(0);
313: }
317: PetscErrorCode EPSView_BLZPACK(EPS eps,PetscViewer viewer)
318: {
320: EPS_BLZPACK *blz = (EPS_BLZPACK*)eps->data;
321: PetscBool isascii;
324: PetscTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&isascii);
325: if (!isascii) {
326: SETERRQ1(((PetscObject)eps)->comm,1,"Viewer type %s not supported for EPSBLZPACK",((PetscObject)viewer)->type_name);
327: }
328: PetscViewerASCIIPrintf(viewer," BLZPACK: block size=%d\n",blz->block_size);
329: PetscViewerASCIIPrintf(viewer," BLZPACK: maximum number of steps per run=%d\n",blz->nsteps);
330: if (blz->slice) {
331: PetscViewerASCIIPrintf(viewer," BLZPACK: computational interval [%f,%f]\n",eps->inta,eps->intb);
332: }
333: return(0);
334: }
338: PetscErrorCode EPSSetFromOptions_BLZPACK(EPS eps)
339: {
341: EPS_BLZPACK *blz = (EPS_BLZPACK *)eps->data;
342: PetscInt bs,n;
343: PetscBool flg;
346: PetscOptionsHead("EPS BLZPACK Options");
348: bs = blz->block_size;
349: PetscOptionsInt("-eps_blzpack_block_size","Block size","EPSBlzpackSetBlockSize",bs,&bs,&flg);
350: if (flg) {EPSBlzpackSetBlockSize(eps,bs);}
352: n = blz->nsteps;
353: PetscOptionsInt("-eps_blzpack_nsteps","Number of steps","EPSBlzpackSetNSteps",n,&n,&flg);
354: if (flg) {EPSBlzpackSetNSteps(eps,n);}
356: PetscOptionsTail();
357: return(0);
358: }
360: EXTERN_C_BEGIN
363: PetscErrorCode EPSBlzpackSetBlockSize_BLZPACK(EPS eps,PetscInt bs)
364: {
365: EPS_BLZPACK *blz = (EPS_BLZPACK*)eps->data;;
368: if (bs == PETSC_DEFAULT) blz->block_size = 3;
369: else if (bs <= 0) {
370: SETERRQ(((PetscObject)eps)->comm,1,"Incorrect block size");
371: } else blz->block_size = PetscBLASIntCast(bs);
372: return(0);
373: }
374: EXTERN_C_END
378: /*@
379: EPSBlzpackSetBlockSize - Sets the block size for the BLZPACK package.
381: Collective on EPS
383: Input Parameters:
384: + eps - the eigenproblem solver context
385: - bs - block size
387: Options Database Key:
388: . -eps_blzpack_block_size - Sets the value of the block size
390: Level: advanced
391: @*/
392: PetscErrorCode EPSBlzpackSetBlockSize(EPS eps,PetscInt bs)
393: {
399: PetscTryMethod(eps,"EPSBlzpackSetBlockSize_C",(EPS,PetscInt),(eps,bs));
400: return(0);
401: }
403: EXTERN_C_BEGIN
406: PetscErrorCode EPSBlzpackSetNSteps_BLZPACK(EPS eps,PetscInt nsteps)
407: {
408: EPS_BLZPACK *blz = (EPS_BLZPACK*)eps->data;
411: if (nsteps == PETSC_DEFAULT) blz->nsteps = 0;
412: else { blz->nsteps = PetscBLASIntCast(nsteps); }
413: return(0);
414: }
415: EXTERN_C_END
419: /*@
420: EPSBlzpackSetNSteps - Sets the maximum number of steps per run for the BLZPACK
421: package.
423: Collective on EPS
425: Input Parameters:
426: + eps - the eigenproblem solver context
427: - nsteps - maximum number of steps
429: Options Database Key:
430: . -eps_blzpack_nsteps - Sets the maximum number of steps per run
432: Level: advanced
434: @*/
435: PetscErrorCode EPSBlzpackSetNSteps(EPS eps,PetscInt nsteps)
436: {
442: PetscTryMethod(eps,"EPSBlzpackSetNSteps_C",(EPS,PetscInt),(eps,nsteps));
443: return(0);
444: }
446: EXTERN_C_BEGIN
449: PetscErrorCode EPSCreate_BLZPACK(EPS eps)
450: {
452: EPS_BLZPACK *blzpack;
455: PetscNewLog(eps,EPS_BLZPACK,&blzpack);
456: eps->data = (void*)blzpack;
457: eps->ops->setup = EPSSetUp_BLZPACK;
458: eps->ops->setfromoptions = EPSSetFromOptions_BLZPACK;
459: eps->ops->destroy = EPSDestroy_BLZPACK;
460: eps->ops->reset = EPSReset_BLZPACK;
461: eps->ops->view = EPSView_BLZPACK;
462: eps->ops->backtransform = EPSBackTransform_BLZPACK;
463: eps->ops->computevectors = EPSComputeVectors_Default;
465: blzpack->block_size = 3;
466: blzpack->slice = 0;
467: blzpack->nsteps = 0;
469: PetscObjectComposeFunctionDynamic((PetscObject)eps,"EPSBlzpackSetBlockSize_C","EPSBlzpackSetBlockSize_BLZPACK",EPSBlzpackSetBlockSize_BLZPACK);
470: PetscObjectComposeFunctionDynamic((PetscObject)eps,"EPSBlzpackSetNSteps_C","EPSBlzpackSetNSteps_BLZPACK",EPSBlzpackSetNSteps_BLZPACK);
471: return(0);
472: }
473: EXTERN_C_END